Designing a multi-classifier system
Conventional crop classification methods do not provide the accuracy required for large-scale crop mapping of smallholder farms in sub-Saharan Africa due to complex crop-planting patterns (including mixed cropping), environmental and field management heterogeneity, small fields and vaguely delineated field boundaries.
In general, the lack of accuracy in image classification has been addressed by adding more data or by improving the efficiency of the algorithm. The former is typically done by including vegetation indices, textural features or by working with image time series. The latter is addressed using advanced classification algorithms or by combining existing ones.
In my research project, I developed a supervised multi-classifier system that combines different machine learning classification methods to produce more accurate crop maps. This system was built in the cloud and, as such, it could manage an expanded set of multi-temporal features like vegetation indices and texture.
Multispectral bands were combined with the panchromatic image to give us pan-sharpened images. This allowed to retain both spatial and spectral resolution. Below, we show a comparison between two subsets of these images. We can see (in the square box) that pan-sharpened image presents more details of crops.
- Figure 2. Subsets of original multi-spectral image and its corresponding pan-sharpened image, the latter one providing more details than the multi-spectral image.
A set of experiments were executed to evaluate the performance of the multi-classifier system and results showed that, indeed, the multi-classifier outperformed individual classifiers. More precisely, the multi-classifier led to an accuracy improvement of up to 21,9% compared to the accuracy of the individual classifiers. We can observe in the figure below the comparison between kappa values of individual classifiers and the kappa values of multi-classifiers. Weighted methods seem to do better than majority voting, but the choice of weighting scheme is not so fundamental.
Figure 3. Accuracy of single classifiers (blue - far left) and multi-classifier (other colours). Combination methods: majority voting and weighted majority voting with different weights (kappa, fscore, fscore per class and producer accuracy per class).
We may notice that kappa values that my study produced are low to be workable, and we need to do further research to explain it since much can be learned in this domain.
Nevertheless, this work reveals the potential of multi-classifiers in addressing complex problems where single classification approaches are not sufficiently successful. Clearly, the use of a multi-classifier in the case here results in a decent quality improvement in classification.